Integrated Learnware Identification and Reuse via Reusability-Aware Metric Learning

The learnware paradigm aims to help users solve machine learning tasks by reusing existing well-trained models instead of building from scratch. These models are accommodated in a learnware dock system, where each learnware consists of a model and a specification that characterizes the model's capability, enabling the learnware to be effectively identified and reused for future tasks. Based on specifications, the system identifies learnwares whose training distributions align with the user task. However, whether the identified learnwares benefit the user depends not only on task similarity but also on how they perform after reuse, making it important that identification reflects expected reuse performance. To this end, this paper leverages the reuse experience among hosted learnwares to make specifications characterize learnware reusability for new user tasks. Specifically, the system treats each hosted learnware as a pseudo-user task and evaluates the task performance achieved by identifying and reusing other learnwares for these tasks. Under the supervision of these system-wide evaluations, the system evolves specifications into a reusability-aware specification space where shorter distances correspond to higher expected reuse performance. With this specification evolvement, learnware identification is guided by expected reuse performance rather than original task similarity alone. Experiments demonstrate that, with reusability-aware identification, the system can effectively recommend learnwares that benefit new user tasks after reuse.